Counterfactual Training

Status

• Code base: In place and streamlined for reproducibility and configuration.
• Experiments: Lots of work done and results largely supportive of idea.
  • Ran into problems on DelftBlue, which has set me back about 2 weeks.
• Paper: Still bare-bones.
• ICML: Potentially still possible to submit something, but this will be rushed and not “finished”.

High-Level Idea

Combine ideas from Energy-Based Models and Adversarial Training:

\[ \ell_{\text{clf}}(f_\theta(x),y) + \lambda_{\text{gen}} \ell_{\text{gen}}({x^\prime}_t,x_t;\theta) + \lambda_{\text{adv}} \ell_{\text{clf}}(f_\theta({x^\prime}_t),y) \]

• \({x^\prime}_t\) are counterfactuals of \(x_{s}\subseteq x\) with target class \(t\).
• \(\ell_{\text{gen}}\) is the difference in energies between observed samples and counterfactuals.
• Counterfactuals are recycled as adversarial examples.

Motivation and Intuition

• Instead of using SGLD to sample from \(p(x|t;\theta)\), we use counterfactual generators.
• The idea is to align counterfactual explanations with observed data to induce plausibility.
• This should only work if counterfactuals are generated faithfully (favorable evidence).
• Approach can be leveraged to impicitly encode mutability and domain constraints in model.

Encoding Domain Knowledge

Example: Let \(f_\theta(x)=\theta^{T}x\) be a linear classifier. Suppose that featue \(x_1\) is immutable (e.g. ‘gender’ or ‘faith’).

\[ \begin{aligned} \nabla_\theta \ell_{\text{gen}}(x^\prime,x;\theta) &= \nabla_\theta ( \theta^{T}x[t] - \theta^{T}x^\prime[t]) \\ \frac{\partial \ell_{\text{gen}}}{\partial\theta_1}(x^\prime,x;\theta)&=x_1 - x_1^\prime \end{aligned} \]

• If $$